PyData Amsterdam 2026

Simulating real-world human institutions with LLMs introduces unique technical challenges.
For example, how do you enforce strict procedural rules without destroying dynamic debate
flow? How do you keep agents grounded in distinct political identities without leaning into
caricature? And how do you measure the systemic impact of structural changes? These
topics will be addressed in this presentation.

Outline:

1. Introduction & Problem Statement (4 min): Context on parliament simulation,
   project scope, and the motivation behind modeling democratic reforms.
2. Architecture & Deterministic State (6 min): Using LangGraph to manage state
   transitions, turn-taking, and procedural parliamentary rules throughout the debate.
3. Traceability & Debugging (3 min): Using LangSmith to monitor prompt inputs, tool
   calls, and token costs for debugging
4. Grounding Agents & Mitigating Bias (6 min): Combining structured context,
   party-filtered RAG, and voting histories to ground agent speeches and reduce base
   model bias.
5. A/B Testing Systemic Reforms & Findings (7 min): Evaluates structural
   democratic reforms by simulating 238 debate runs to measure their impact on debate
   quality and policy outcomes.
6. Limitations & Conclusions (4 min): Addressing the inherent limitations of using
   LLMs for simulations and validating transcripts with domain experts.
7. Q&A

Takeaways:

• How do you transfer a real-world system such as debates into a full multi-agent
  simulation
• How to control next-turn decisions in this multi-agent using custom bidding phases
• Grounding the agent personas using RAG to make sure the agent doesn’t rely on
  internal knowledge
• How to create an A/B experiment to test systemic changes.

Impact

The impact of this project created additional evidence that reforms in the parliament are
necessary. The findings have been used in a report by this think tank with advice on how to
strengthen democracy and is read by politicians, journalists and policy experts.
Generalized applications
This multi-agent simulation could be interesting in any team environment where you would
like to have multiple agents with distinct opinions, expertise and philosophies. You could
instruct them to come to a consensus and see how that plays out. For instance, engineering
teams, executive boards, focus groups or cross-functional teams could be simulated.

Generalized applications

This multi-agent simulation could be interesting in any team environment where you would
like to have multiple agents with distinct opinions, expertise and philosophies. You could
instruct them to come to a consensus and see how that plays out. For instance, engineering
teams, executive boards, focus groups or cross-functional teams could be simulated.

Target audience

This talk is interesting for Machine Learning Engineers, Data Scientists, and Developers who
want to start or have experience with building AI agents, RAG and layered prompting